AI stocks after the sell-off: why the long-term investment case may remain intact

Markets have been unsettled since late June, with AI infrastructure, semiconductors and memory among the areas affected. In this Q&A, Portfolio Manager Yash Patodia looks at what may be behind the move, why the long-term AI investment case may still be supported, and where select opportunities could be emerging across the AI value chain.

AI stock sell-off: what is really driving the volatility?

In our view, we would characterize the recent weakness as a positioning-driven unwind more than a broad deterioration in the long-term AI investment case. AI-related stocks had become crowded across a range of investor strategies, which left the group vulnerable to rapid selling when sentiment shifted. As investors reduced risk, pressure spread across AI infrastructure, semiconductors and memory, with the impact particularly visible in parts of Asia.

Importantly, share-price weakness alone does not tell the full story. Some of the selling appears technical and broad-based rather than clearly tied to company-specific fundamentals. That said, volatility could persist. Investors should be prepared for further swings, even if the underlying demand picture remains supportive.

Are concerns about AI infrastructure spending, open-source models and China memory supply justified?

Investors are right to ask harder questions after such a strong run. We see three main areas that warrant attention. First, some large AI infrastructure projects appear to involve complex financing arrangements, including vendor-backed support. These structures deserve scrutiny because tighter financing conditions, weaker funding availability or changes in market confidence could affect AI-related equities even if end demand remains healthy.

Second, some investors worry that more efficient, lower-cost open-source AI models could reduce the need for expensive hardware. We think the relationship is more nuanced. Lower-cost AI can make adoption easier and broader, which may increase total demand for computing power over time. If AI becomes cheaper and easier to deploy, more companies and consumers may find use cases that justify additional compute.

Third, concerns around Chinese memory supply have resurfaced. This is an important development to monitor, particularly for conventional memory markets over the medium term. Based on the evidence currently available to us, however, we do not see clear signs of a near-term supply surge in the most advanced AI memory segments. In our view, the immediate constraint in high-bandwidth memory remains meaningful.

Have AI fundamentals weakened — or is the long-term growth story still intact?

So far, in our view, the fundamental picture remains broadly supportive, and in some areas it has improved. Across the AI ecosystem, demand indicators remain healthy, cloud providers continue to invest, and many companies still describe themselves as constrained by available computing capacity rather than by a lack of customer demand.

Recent reporting across the ecosystem supports that view, although the picture is not uniform across companies or segments. Among large cloud platforms, consumer-facing AI usage has continued to scale, cloud revenue growth has generally remained firm, backlogs remain elevated, and capital spending plans have not shown broad-based retrenchment. These signals suggest that capacity remains tight and that many companies continue to see strategic value in AI infrastructure. At the model layer, improvements in compute efficiency may also support better inference economics, although the pace and scale of monetisation will remain important to monitor.

AI’s memory bottleneck: what investors should know about supply constraints

We see memory as an increasingly important bottleneck in the AI build-out. Advanced AI systems require not only more processors, but also high-bandwidth memory, or HBM, to move and process large amounts of data quickly. As demand for AI compute grows, HBM has become a key constraint across parts of the semiconductor value chain.

We believe this supply constraint is real and could persist for some time. Adding advanced memory capacity is difficult and time-consuming. New cleanroom space can take years to build, and based on our current assessment, limited meaningful new capacity may become available across major producers before late 2027 or early 2028. A broader supply-and-demand balance may take longer to emerge, although timing remains uncertain and subject to change.

The key point is that HBM does not simply add a new product category to the memory market — it also uses scarce manufacturing capacity. Producing HBM requires materially more wafer capacity per bit than conventional DRAM, and that capacity burden tends to increase with each new generation. As producers allocate more resources to HBM, supply in conventional DRAM can also tighten. This is why higher capital spending does not necessarily translate one-for-one into more usable memory supply. Memory remains cyclical, and a correction at some point would be normal. However, this cycle has some distinguishing features: demand is increasingly linked to AI infrastructure spending, the supply response is constrained by physical capacity, and longer-term customer agreements could support greater durability than in prior cycles. Investors should continue to watch for signs of new supply, softening demand or inventory accumulation.

Are low AI semiconductor valuations a warning sign — or an opportunity?

This is the right question to ask, and our discipline is to underwrite normalised rather than current returns. Low valuations can sometimes signal that the market believes earnings are close to a peak. However, in parts of the AI semiconductor complex, earnings revisions have moved up alongside — and in some cases faster than — share prices. This suggests that low multiples should not be dismissed automatically, but they need to be assessed carefully.

For investors, the key question is whether today’s earnings power is sustainable or simply reflects unusually tight supply and elevated pricing. A low multiple on peak earnings can become a value trap if profits normalise quickly. But if AI demand proves more durable and supply remains constrained for longer, current valuations may not fully reflect the potential quality and duration of the earnings cycle. This is why selectivity matters. We would focus less on headline valuation alone and more on where fundamentals, cash flows and competitive positioning appear resilient through the cycle.

Where are the next AI investment opportunities across the semiconductor value chain?

In our view, the bottlenecks now extend beyond the chips themselves. We see potential opportunities emerging across substrates, test equipment, optical components, storage and power infrastructure — areas that are increasingly important to scaling AI systems. The distinction we make is between memory volume growth, which we believe can remain robust, and memory pricing, which is inherently harder to forecast.

That has drawn our attention to businesses tied to long-term capacity expansion rather than short-term pricing alone. Areas of interest include semiconductor capital equipment, where each new generation of HBM raises test intensity; NAND and storage, where the industry appears underbuilt relative to potential demand; and advanced packaging, substrates and testing, which are increasingly important constraints on system output. At the same time, moving further down the value chain is not automatically safer or cheaper. Some second-order beneficiaries already trade at premium valuations, while AI-related cost inflation can pressure areas such as PCs, smartphones and consumer electronics. In our view, the opportunity set is broadening, but selectivity and valuation discipline are becoming more important.

What should investors watch next after the AI sell-off?

For us, the more important signal would be a change in the fundamental evidence, rather than share-price volatility alone. Key warning signs would include AI adoption failing to broaden beyond early use cases, slower enterprise commercialisation, weaker evidence of customer return on investment, or signs that capacity additions are beginning to outpace demand. The critical long-term indicator, in our view, is not capital spending in isolation, but whether enterprises continue to derive measurable productivity benefits from AI.

Absent such evidence, we would not view the correction as a reason to make wholesale changes. Instead, it may create an opportunity to reassess positioning, rebalance exposures and selectively lean into areas where fundamentals, valuation and competitive positioning remain attractive. With AI adoption still developing — and with potential longer-term applications across agentic AI, robotics and next-generation devices — rising dispersion may create an environment in which active management can add value.